I gave my coding agent a paintbrush. Ask Claude Code or Codex for a poster, logo, app screen or dashboard. It designs it in SVG, and you watch it appear live in your browser. No API keys. No image model. Runs on the plan you already have. Free and open source. Link below.

This is peak cope. Nope, you just lost the edge you had to a lil kid. Software just started and everyone can play now. Industrial Revolution 2.0, read some history kids. And China will burst the bubble, by next 2 years… so it will become free to make any software

Became a dad last night. Meet Cyrus!

Zeus chilling

ok I'm actually so hyped rn 😭 my first ever plugin just got accepted into Claude's plugin directory!! it's called Playcut Brush. you ask Claude to design something (a poster, a logo, an app screen) and you watch it paint it live in your browser totally free, no API keys. go try it 🎨

Made a new friend.

What genre is this song? @grok open.spotify.com/track/0NbME…

Install in Claude Code: /plugin marketplace add mehrang0/playcut-brush /plugin install brush@playcut github.com/mehrang0/playcut-…

sinatra.dev/ really cool project if you haven't seen it yet! @Sinatradev

To use AI fully you need to think recursively only.

Elbow is all you need. :/

Day 5 of building Zeus, a 7B coding model that runs on your laptop. We trained Zeus-D yesterday. Honest results: 🟢 Fixing real GitHub bugs: 68% → 84%, with a third fewer tokens 🔴 Contest puzzles: 52 → 43. It learned to overthink and ran out of room. We wrote the pass/fail rule before training. By that rule, it failed. So it doesn't ship. Next: keep the coding gain, fix the overthinking, retrain.

Day 4 of building Zeus One GPU. One epoch. 7 skills. Zeus-D is learning: loss on held-out tasks is down across the board at the 25% mark. The real test comes after: does it beat Zeus-C inside a coding agent? We wrote the win rule before training so we can't fool ourselves.

Zeus-C update: our 7B coding model went 38.3% → 52.0% on LiveCodeBench. LiveCodeBench ranking: 1. DeepSeek V4.1 Flash: 92.5% 2. Claude Opus 4.8: 87.8% 3. Claude Opus 5: 74.6% 4. Zeus-C (7B): 52.0% 5. K2-Horizon-7B base: 38.3% Hard problems: 7 → 25 of 80. No drop on HumanEval+/MBPP+. Runs fully local on a 2021 MacBook Pro: M1 Pro · 16 GB · 4-bit (5.6 GB) 22.7 tok/s writing · 235 tok/s reading

Building Zeus in public: an open-source 7B coding model that runs fully on a laptop. No cloud. Goal: DeepSeek-level coding on a 16 GB MacBook. Day 3. Day 1: pick the base Our 30-task coding set, run on a 2021 M1 Pro: • Qwen3.5-9B: 3/13 • MiniCPM5-2B: 8/30 • K2-Horizon-7B: 30/30 at 22.7 tok/s K2 won. We built our own agent harness around it: plan tree, tools, and tests as the judge. It went 30/30 inside the loop too. Day 2: first failure, then a teacher Training the model on its own best runs failed 3 times, because it already predicts itself. So we switched to DeepSeek as the teacher, keeping only the runs our tests verify. → 960 training examples for $1.23 Day 3: Zeus-B, trained • 22% lower loss on tasks it never saw • HumanEval+ 84.1 / MBPP+ 75.4, no regression • Agent loop: 29/32 solved vs base 26/31 • Fewer tokens on every task Day 3: Zeus-C, the hard problems • 5,315 Codeforces problems • 995 solutions, each verified on hidden tests • Ratings up to 2800 • Up to 30k tokens of reasoning per answer • $14.38 of teacher calls Training tonight. LiveCodeBench next. Rented GPUs for a few dollars. The model runs on a laptop. Every number here comes from a test we ran.

Making a model named after my dog lol

everyone's hyped on jev. read the blog, the docs, and the one hands-on review. here's what's actually there and what isn't 1/ what it is: a classifier. you give it a doc + a list of options, it returns probabilities in one forward pass. no text generation. that's the whole trick. classifiers have done this for a decade, the new part is the training target (calibration instead of preference) and the packaging 2/ "194x faster, 445x cheaper". against what? the average of gpt-6 astra and claude fable thinking for 8 seconds. nobody uses a reasoning model to route a support ticket. the fair baseline is a fine-tuned 0.5b classifier or reading logprobs off an open model. both near free, both near instant. that comparison isn't published 3/ accuracy: zero numbers. no f1, no ground truth benchmark, no calibration curve. the blog literally says "we don't optimize for a ground truth classification". the only independent test (every.to) had it catch 6 of 7 planted defects. the llm it's supposed to replace caught 7 of 7 4/ "0% hallucination". it can't write a sentence, so it can't write a false one. that's a definition, not a result. it can still be confidently wrong, and calibrated confidence just means it's wrong at the advertised rate 5/ no reasoning. one pass is system one by design. can't chain steps, can't look anything up, can't run a check, can't ask a question back. the decisions that matter are exactly the ones that need that 6/ no action. it can't fix the invoice, write the reply, or produce the code. every workflow still needs an llm or a real engine next to it. "composable intelligence" means you compose it 7/ the actual miss: a decision model is the one model that sees its own outcomes. did the ticket get resolved, was the invoice actually fraud. that's free labels every day. nothing in the blog or docs says jev learns from any of it. it ships frozen like everything else, and that was the one thing it could have done that an llm can't 8/ closed. no weights, no size, no context length, no on-device. every decision goes through their api at $0.042/M in. cheap until it's your whole pipeline 9/ what's real: calibration as a training objective (rlcd) is a good idea, and a cheap fast judge is useful. i'd use it as a gate in front of a real engine or a rubric scorer for drafts. that's a feature, not a new kind of intelligence 10/ the two things that would make it matter: publish accuracy + ece against cheap baselines, and recalibrate from prod outcomes. until then it's a fast classifier with good marketing from someone who earned the benefit of the doubt